Hybrid Prediction Model Of Soybean Futures Price Based On Machine Learning

Soybean is an important food and cash crop in China.With the rapid development of China's food processing and animal husbandry,people's demand for processed soybean products such as edible soybean oil,vegetable protein and feed soybean meal is increasing,and China has become a major importer of soybean trade.Soybean prices have frequently fluctuated due to the comprehensive influence of international markets,politics and economy,which has brought new problems and challenges to China's food security,soybean production and industry benefits.Effective prediction of the transaction price of the soybean futures market can help China's soybean enterprises or farmers reasonably avoid the risk of market price fluctuations,control the planting structure to ensure market circulation,and participate in futures trading hedges,thereby ensuring China's food security,industry interests,and industrial stability.The data in this article is derived from the Dalian Soybean No.1 soybean futures data,and the samples obtained for 20,10,3 and 1 years are 4812,2398,698 and 209 respectively.Based on the introduction of traditional time series differential autoregressive moving average(ARIMA),exponential smoothing,machine learning(ML)and their mathematical models,by selecting the sample data whose autocorrelation coefficient is above 0.70 as the input variable,and the current index price as the output variable,using support vector regression(SVR),gradient boosting regression tree(GBRT)and long-term short-term memory network(LSTM)models for prediction.This paper builds a mixed prediction model of soybean futures price based on machine learning,which can effectively extract the linear and nonlinear information of financial data and improve the accuracy of soybean futures forecast.The model is based on the mixed model of ARIMA?ML and ES?ML,and the residuals of the predicted and true values obtained by the ARIMA and ES models are used as the new time series,and then the residuals are fitted by SVR,GBRT,LSTM,and the sum of the two is used as the final prediction result.After empirical analysis and comparison,ARIMA?ML model is better than ARIMA,and ML is better than ARIMA?ML model;ES?ML model is better than ES model,the mixed model has a good prediction effecton short-term data that meets the normal distribution and has little fluctuation,the optimal MAPE of the ES?LSTM model is 0.21%,and the RMSE is 8.23,machine learning models have better prediction effects on long-term data with multi-peak distribution and large fluctuations.The conclusions obtained show that the machine learning model is better than the traditional time-domain prediction model,but the hybrid model is not necessarily better than the machine learning model,and only a suitable hybrid model can effectively improve the prediction accuracy.